Research On High Dynamic Range Inverse Tone Mapping Algorithm Based On Generative Adversarial Network

With the development of high-definition-television and HDR(high dynamic range)technology,the demands on HDR images and videos are increasing rapidly these years.But producing HDR contents directly costs a lot,we need an inverse tone mapping algorithm to transfer widely used SDR(standard dynamic range)to HDR,which could improve visual effect.Regarding to existing inverse tone mapping algorithm,they hardly manage the circumstances with various scenes and do not perform well due to lack of non-linearity.This means there are a lot of room for improvement.The methods of building and editing HDR videos and images are different from ordinary SDR.The standard of HDR recommend that the video and image are quantized in 10 bits with PQ or HLG EOTF/OETF.This means special processing methods for HDR contents.In this paper,when building the dataset,we applied several degrading methods for high quality HDR10 videos to get the corresponding SDR,due to the lack of high quality HDR and application of our algorithm.The dataset is taking the leading position in quantity and quality comparing to other researches.For the first time,we propose a generative adversarial network based algorithm to transfer input SDR image to high quality HDR,focusing on improving the quality of HDR generated.The algorithm reconstructs details of real scene with adversarial training of GAN.The model was designed for large resolution images.The training process based on a hybrid objective function with MSE and differential MSE.Experiment results show that the proposed algorithm generates HDR images much closer to real ones than that generated by existing algorithms both subjectively and objectively.Meanwhile,we perform several experiments to prove the robustness and parameter sensitivity of proposed inverse tone mapping algorithm.Besides,for the first time,we proposed an algorithm which can restore the unsuitable exposure of input SDR images and mapping them to high quality HDR,in allusion to the circumstances that input images are often with poor exposure or brightness.Based on the proposed inverse tone mapping algorithm before,the improvements are adding pre-processing histogram equalization,intrinsic image decomposition and perceptual loss for objective function for strict constraint on intrinsic brightness and color.After adversarial training,we get a network that can restore the exposure to the normal range.Through experiments on testing dataset,it proves that the network generates high quality HDR images from poor SDR images.Meanwhile,in order to prove the robustness of the model,we performed several experiments to prove the necessity of histogram equalization,intrinsic image decomposition and perceptual loss.Then,we test our algorithm on photos of real scenes,and the results showed that the model is effective for practical application.